The centroid shifts with finger angle and pressure
Aliases: centroid shift · finger angle · touch offset
What it is
The previous card established that the system must reduce contact to a point; this one is about how that point itself moves systematically with finger posture. A steeper contact angle elongates the patch toward one side and shifts the centroid with it; more pressure enlarges the contact area and conductive distribution and moves the centroid again. This is not random noise — it is a repeatable, predictable, and in principle correctable bias, which is exactly what "touch offset" refers to in the literature.
Why it happens
Why does angle or pressure move the centroid in a consistent direction rather than just enlarging the error? A finger first contacts the screen with the pad, and as pressure increases the contact region expands toward the base of the finger because skin and pulp deform under load; the finger's tilt determines which side that expansion favours — a tilted finger typically elongates the patch in the direction of the tilt, pulling the centroid away from the point directly beneath the fingertip. Work on modelling touch offset has shown this shift correlates with finger pitch and with position on the screen, and is repeatable for a given person — meaning it is predictable rather than purely stochastic. If the offset were pure noise, no statistical model could remove it; because it is driven by geometry, a correction function taking angle, pressure, and position as input can cancel most of it. That predictability is the entire premise behind later offset-compensation techniques.
Studying it
Studies control pressing angle (near-vertical versus tilted by a set amount), force (light/medium/firm), and finger (thumb versus index), then measure the reported point's displacement and spread relative to a physical reference target. The key methodological point is that the offset must be decomposed by direction, not summarised as a single average error in millimetres: if the bias is systematic but points in opposite directions for different people, averaging can cancel the direction out and make the error look small, while every individual is in fact reliably off in one consistent direction. A correction learned under one posture or device also cannot be assumed to transfer to another — individual variation is large, and cross-condition compensation models generally need recalibration.
Where it stops holding
Not every touch controller reports angle- or pressure-related signals reliably enough to model: many consumer capacitive screens output only the reduced single-point coordinate, exposing neither contact shape nor pressure, so the application has no raw data to build a compensation model from even if it wanted to. Even where a model exists, a correction trained for one user's grip can fail — or actively make things worse — after a hand swap (left to right), a device change, or a posture change (lying down versus sitting), because the direction of the underlying bias itself flips. This is largely why most products settle on enlarging tolerance rather than precisely compensating for offset: compensation only pays off when posture is stable, and that assumption is fragile in real use.
Applying it
- Do not assume the reported touch coordinate equals the finger's geometric centre; hit-tolerance design should default to a directional bias, not just an isotropic fuzzy radius.
- Only invest in offset compensation where raw contact data (area, orientation) is available and the use posture is relatively fixed, such as a tablet held flat and used with both hands; one-handed mobile use varies posture too much for compensation to pay off.
- Revalidate any compensation model after a hand or device change rather than assuming a model trained once will generalise.